/alibabacloud-agent-identity-agentrun-e2e

name: alibabacloud-agent-identity-agentrun-e2e
description: >
  Deploy and test an Alibaba Cloud agent with end-user authentication and
  tool authorization, then clean up test resources. Use when someone wants
  to deploy an AgentRun-hosted agent governed by AgentIdentity and verify
  that only signed-in users reach it, that each tool receives the
  credentials it needs without any secret in the code, and that authorization
  rules allow or deny expected operations — or when they want to remove the
  resources such a trial created. Covers first-time setup on an empty account,
  re-verification after a change, and cleanup.
  也响应"AgentRun 端到端测试""AgentIdentity 全流程验证""AgentRun 部署 Agent"等中文请求。
license: Apache-2.0
compatibility: >
  Requires macOS or Linux (zsh/bash), Python >= 3.10, conda (miniforge or
  miniconda), Alibaba Cloud CLI (aliyun) >= 3.3.3 — older versions lack the
  plugin-mode commands this skill calls; upgrade with `brew upgrade aliyun-cli`
  or from https://github.com/aliyun/aliyun-cli/releases. Also needs an Alibaba
  Cloud account with AgentRun, AgentIdentity, RAM and OSS permissions (see
  references/ram-policies.md) and network access to Alibaba Cloud endpoints.
metadata:
  version: "1.0.0"
  layer: application
  category: deployment
  lifecycle: stable
  tags: "agentrun,agent-identity,end-to-end,deployment,testing"

Alibaba Cloud AgentRun + AgentIdentity End-to-End Deployment & Test

Orchestrates the full e2e workflow: CLI credential setup → identity provider registration → OAuth2 credential chain → console MCP registration → sample build & Runtime deployment → inbound/WAT/Cedar verification → local-tool credential injection → DingTalk MCP → cleanup.

The runtime code lives in the [agent-identity-dev-kit](https://github.com/aliyun/agent-identity-dev-kit) repository under `agent_identity_python_samples/`. The prerequisites and phases below describe the additional inputs and setup required.

Execution Rules

Run shell blocks in Bash from the Skill directory. Supply the same absolute `TMPDIR` on every invocation so each script reads the same run state.

1. Blocks marked `<!-- AGENT:AUTO -->` — execute directly in the terminal, no user confirmation needed. 2. Blocks marked `<!-- AGENT:WAIT:description -->` — STOP and wait for the user's response before proceeding. 3. CLI calls use the configured aliyun profile; Python SDKs use their default credential chain. Ensure both resolve to the intended account; credential environment overrides can change the SDK identity. If the CLI is not configured, guide the user through `aliyun configure`; verify with `aliyun sts get-caller-identity` before cloud preparation. Do not use a separate secrets file or pass, print or export an AccessKey pair. 4. On command failure, stop and report the error (with logs) to the user. Do NOT retry automatically; retry only after the user confirms. Match errors against `references/troubleshooting.md` first. 5. Terminal sessions lose exported variables between calls. Durable state lives in files under the skill state directory; scripts re-read them on every run. All scripts source `scripts/lib/common.sh`. 6. Follow the phases in order; do NOT skip, merge, or simplify steps. Present each WAIT step to the user exactly as written. 7. Background services (nohup) are started and stopped INSIDE the scripts — never append `&` to a terminal call yourself. 8. Automate everything that can be automated — including cloning the kit repository and installing missing tools — and tell the user in one line what you did. Ask the user to act ONLY where their person is strictly required: entering secrets, console UI steps, and account-level decisions. 9. Environment variables in deployment templates are minimal: required variables only. Never add debug switches (e.g. AGENTRUN_SDK_DEBUG) or empty placeholders (e.g. OAUTH_SCOPES=""). 10. Register MCP tools through the AgentRun console. This workflow uses the console registration path verified for tool activation and AgentIdentity configuration. 11. OAuth2 authorization links expire within minutes. Whenever a link is returned: (a) IMMEDIATELY open it in the user's browser yourself — run `open "<url>"` on macOS or `xdg-open "<url>"` on Linux — do not rely on the user noticing the chat message in time; (b) also print the link in the message as a fallback. Access-token expiry depends on the provider configuration, so later runs may require authorization again. 12. Resource naming is the USER's decision. At every resource-creation step, either ask the user for the name (offer a suggested default) or, when the name is fixed by the sample code (e.g. the `test-provider-api-key` credential provider), state the required name and the reason explicitly. Never silently pick names. More generally: before each step, tell the user what it will create/change and why — the user must always know what is being done on their account. 13. Every message that starts, advances, blocks, or closes a phase follows the mandatory shape in "Output Format" below. Dropping the progress header or the two sections on such a message is a defect, even when the content is correct. 14. All aliyun CLI calls use plugin mode: kebab-case commands and flags (`aliyun sts get-caller-identity`, `aliyun ram attach-policy-to-role --policy-type ...`). Phase 1 enables `auto-plugin-install` and pre-installs the needed plugins; without that, a missing plugin makes the CLI prompt for confirmation and then abort on `failed to read user input: EOF`. 15. Every cloud call carries the skill's User-Agent (see "Observability" below). Do not strip it or issue raw calls that bypass it.

Output Format

Every message that starts, advances, blocks, or closes a phase MUST use the shape below. That is the hard requirement, and it covers the whole main path.

For a reply that merely reacts inside an already-running flow — a how-to question, a failure report, a refusal — keep the same shape whenever the flow is underway, so the use